Machine Status Prediction for Dynamic andHeterogenous Cloud Environment

Jinliang Xu, Ao Zhou∗, Shangguang Wang, Qibo Sun, Jinglin Li, Fangchun YangState Key Laboratory of Networking and Switching Technology

Beijing University of Posts and Telecommunications, Beijing, China

{jlxu,aozhou,sgwang,qbsun,jlli,fcyang}@bupt.edu.cn

Abstract—The widespread utilization of cloud computing ser-vices has brought in the emergence of cloud service reliabilityas an important issue for both cloud providers and users. Toenhance cloud service reliability and reduce the subsequent losses,the future status of virtual machines should be monitored in realtime and predicted before they crash. However, most existingmethods ignore the following two characteristics of actual cloudenvironment, and will result in bad performance of status pre-diction: 1. cloud environment is dynamically changing; 2. cloudenvironment consists of many heterogeneous physical and virtualmachines. In this paper, we investigate the predictive power ofcollected data from cloud environment, and propose a simple yetgeneral machine learning model StaP to predict multiple machinestatus. We introduce the motivation, the model developmentand optimization of the proposed StaP. The experimental resultsvalidated the effectiveness of the proposed StaP.

Keywords—cloud environment, status prediction, heterogenous,dynamic.

I. INTRODUCTION

The large-scale utilization of cloud computing services for

hosting industrial applications has led to the emergence of

cloud service reliability as an important issue for both cloud

service providers and users [1]. Cloud environment employs

plenty of physical machines that are interconnected together

to deliver highly reliable cloud computing services. Although

the crash probability of a single virtual machine might be low,

it magnifys across the whole cloud environment [2].

To enhance cloud service reliability and reduce the subse-

quent losses, the future status of virtual machines, such as

missing, locked, paused, deleting, the time taken to crash,

etc., should be predicted before it happens. However, most

existing methods ignore the following characteristics of actual

cloud environment and this will result in bad performance

of status prediction [1], [3]: (1) the cloud environment is

dynamically changing, which requires response in real time;(2)

cloud environment consists of many heterogeneous physical

machines, which bring about different items to deal with,

such as fan speed, CPU temperature, memory, disk error,

application updating, software architecture, etc.

In this paper, to tackle the above problems, we propose a

simple yet general machine learning model StaP to predict

multiple machine status. StaP can automatically learn the

representation of of different items and the correlations among

them, and predict multiple statuses in real time when ready

trained. The development and optimization process of the

proposed StaP is detailed. The experimental results validated

the effectiveness of StaP.

II. THE PROPOSED MODEL

Let V = {�vm ∈ RD|m = 1, 2, · · · ,M} denote representa-

tion vectors of collected items in a D-dimension continuous

space, where vector �vm represent the m-th item bm. V is

shared across all machines and can automatically be learned

from collected data. Let Sn denote the collected item list of

machine un. The length of Sn and its elements may vary

with different machines. Then we denote collected data by

R = {rn,m|n = 1, · · · , N,m ∈ Sn}, where rn,m stand for the

collected value of un on bm. We use one-hot representation

to represent each status, and concatenate them all to generate

representation �yn of all statuses of machine un [4].

Based on these denotations, we can represent the conditional

probability that machine un takes value bm by adding the

corresponding value rm,n as scaling factor as follows:

p(�vm|�yn) = (pr(�vm|�yn))rm,n

=exp

(rm,n�v

TmW �yn

)(∑

m∈Snexp (�vTmW �yn)

)rm,n

, (1)

where W ∈ RD×C is a parameter that need to be learned from

data. And then, according to the product rule in probability

theory, we can get probability that un takes all values of Sn

as follows:

p(Sn|�yn) =∏

m∈Sn

p(�vm|�yn)

=exp

((∑m∈Sn

rm,n�vm)T

W �yn

)(∑

m∈Snexp (�vTmW �yn)

)∑m∈Sn

rm,n

. (2)

Finally we get the objective function as the log likelihood

over all the machines as follows:

�StaP =N∑

n=1

log p(�yn|Sn), (3)

where η is the regularization constant.

2016 IEEE International Conference on Cluster Computing

2168-9253/16 $31.00 © 2016 IEEE

DOI 10.1109/CLUSTER.2016.73

136

Training model is to find the optimal parameters Θ that can

maximize Eq. 3. With elementary algebraic manipulations, we

can change the training target into:

Θ∗ = argmaxW ,V

N∑n=1

{( ∑m∈Sn

rm,n�vm

)T

W �yn

−∑

m∈Sn

rm,n log∑

m∈Sn

exp(�vTmW �yn

)} (4)

As direct optimizing would suffer high computational cost,

we resort to the negative sampling technique [5] for efficiency.

The optimizing process is shown in Algorithm 1, where σ(x)means the logistic function.

Algorithm 1: Learning algorithm

Input: R, Y = {�yn},learning rate η, maximum iterations

maxIt, sampling number k;

Output: W , V ;

1 Initialize W , V ,t = 0, define �ϕn =∑

m∈Snrm,n�vm;

2 while t++ < maxIt do3 for n = 1;n ≤ N ;n++ do4 W = W + ησ

(−�ϕTnW �yn

)�ϕn�y

Tn ;

5 for m ∈ Sn do6 �vm = �vm + ησ

(−�ϕTnW �yn

)rm,nW �yn;

7 for i = 1; i ≤ k; i++ do8 sample negative sample �yi;9 W = W − ησ

(�ϕTnW �yn

)�ϕn�y

Ti ;

10 for m ∈ Sn do11 �vm = �vm − ησ

(�ϕTnW �yn

)rm,nW �yi;

12 Update W = W − 2ληW , V = V − 2ληV ;

13 return W , V ;

With the ready trained parameters W , V , we can predict �yzfor a new machine uz according to

�y ∗z = argmax

�y∈Yp(�y|Sz) = argmax

�y∈Yp̃(�yn)p(Sn|�yn)

= argmax�y∈Y

{log p̃(�y) +

( ∑m∈Sz

rm,z�vm

)T

W �y

−( ∑

m∈Sz

rm,z

)log

( ∑m∈Sz

exp(�vTmW �y

))}, (5)

where p̃(�yn) is the empirical distribution of machine status

representation �yn given by the R.

As the terms(∑

m∈Szrm,z�vm

)TW ,

∑m∈Sz

rm,z and

�vTmW are constant for all �y, the process to get �y ∗z would

not cost much. In Eq. 5, the empirical distribution p̃(�y) can

be considered as the prior probability of �y, and p(Sn|�yn) is

closely related to the likelihood function.

III. EXPERIMENT

The experimental dataset contains 210, 000 ratings ex-

pressed by 1, 075 users on 2, 000 books. A user has individual

information, such as gender, age group and his rating list. We

choose this dataset because a user, his rating list and individual

information can be mapped to a virtual machine, the set of

collected items and its two different future statuses.

We employ POP and SNE as baseline models, and weight-ed F1 and Hamming Loss as evaluation metrics[4]. They

are commonly used in multi-task multi-class classification

problem, which is similar to status prediction tasks in cloud

environment.

TABLE IPERFORMANCE COMPARISON.

Training

ratio (%)

weighted F1 Hamming Loss

POP SNE StaP POP SNE StaP

50 0.095 0.213 0.278 0.464 0.469 0.46770 0.096 0.315 0.350 0.489 0.463 0.45890 0.096 0.367 0.379 0.451 0.452 0.443

The experimental results are as shown in Table I. Clearly,

the proposed StaP outperforms POP and SNE under different

evaluation metrics all the time, as we set the training data ratio

with 50%, 80% and 90% respectively. This result validates the

assumption that the proposed StaP is a more proper model to

predict the current status of virtual or physical machines by

utilizing data that collected from cloud environment.

IV. CONCLUSION

In this paper, we address the problem of virtual machine

status prediction for dynamic and heterogenous cloud environ-

ment. More specifically, we investigate the predictive power of

collected data of different items from cloud environment and

propose a simple yet general machine learning model StaP to

automatically learn the representation of of different items and

the correlations among them, and predict multiple statuses in

real time. The experimental results validated the effectiveness

of the proposed model.

ACKNOWLEDGMENT

This work was supported by NSFC (61272521), NSFC

(61571066, 2016.01-2019.12), and the Fundamental Research

Funds for the Central Universities(2016RC19).

REFERENCES

[1] M. Dong, H. Li, K. Ota, L. T. Yang, and H. Zhu, “Multicloud-basedevacuation services for emergency management,” Cloud Computing,IEEE, vol. 1, no. 4, pp. 50–59, 2014.

[2] P. Gill, N. Jain, and N. Nagappan, “Understanding network failures in datacenters: measurement, analysis, and implications,” in ACM SIGCOMMComputer Communication Review, vol. 41, pp. 350–361, ACM, 2011.

[3] J. Liu, S. Wang, A. Zhou, S. Kumar, F. Yang, and R. Buyya, “Usingproactive fault-tolerance approach to enhance cloud service reliability,”IEEE Transactions on Cloud Computing, pp. 1–13, 2016.

[4] P. Wang, J. Guo, Y. Lan, J. Xu, and X. Cheng, “Your cart tells you:Inferring demographic attributes from purchase data,” in Proceedings ofACM International Conference on Web Search and Data Mining (WSDM),pp. 251–260, ACM, 2016.

[5] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean, “Dis-tributed representations of words and phrases and their compositionality,”in Proceedings of Advances in Neural Information Processing Systems(NIPS), pp. 3111–3119, 2013.

137